1
Artificial Intelligent Research Assistant for Aerospace
Design Synthesis—Solution Logic
Thomas McCall,1 Kiarash Seyed Alavi,2 Loveneesh Rana,3 and Bernd Chudoba.4
AVD Laboratory, UT Arlington Dept. of Mechanical and Aerospace Engineering, Arlington, Tx, 76019, USA
This paper has two objectives. First, the identification and communication of a research
endeavor driving towards an aerospace artificial intelligence design and research assistant.
The second objective is to present the evaluation of a proof of concept to employ a neural
network to aerospace vehicle sizing. This test is part of the development effort to arrive at an
intelligent assistant. The goal is to test the viability of a machine learning augmented synthesis
tool with the intent to decrease the time to design convergence. The topic is approached from
three avenues. First, a consideration of intelligence itself as best defined by humanity is
considered, with the objective to identify key components to intelligence. The consideration of
both natural and artificial intelligence provides direction into the definition of requirements
for the system. This is the second avenue of approach. The second avenue leading to the overall
objective, is the definition of a proposed artificial intelligence design environment. A key
component to this system is identified as a readily modular synthesis approach that is both
quick and computationally attractive. This requirement leads to the development and
subsequent comparison of a neural network synthesis architecture compared to a standard
synthesis software approach—the objective of this document. In order to test and demonstrate
overall functionality, the case study chosen to be implemented is a hypervelocity lifting body
reentry vehicle. The learning sets include five primary design parameters. The final outcome
is the evaluation of a machine learning algorithmic approach to synthesis from a learned
design data set versus the execution results of the subroutine developed synthesis code.
I. Nomenclature
AI = Artificial Intelligence
AVD = Aerospace Vehicle Design
AVD-DBMS = Aerospace Vehicle Design Database Management System
AVD_AI = Aerospace Vehicle Design Artificial Intelligence
Splan = Planform Area
W = load
τ = Slenderness ratio
ΔV = Velocity Required
II. Introduction
ROM the early aerospace vehicle product gestation phase onwards, the future projects engineer is challenged to
develop a level of assurance when committing resources towards a product aimed at achieving the envisioned
impact on the future market years later. The success of a product is dependent on the quality of the underlying early
forecasts. Consequently, the forecasting team or future projects environment is responsible to identify the available
product solution space and risk topographies resulting in the correct choice of the facilitating technologies baseline
design, architecture, or program.
1 PhD Candidate, AVD Laboratory, UT Arlington, Student AIAA Member. 2 PhD. Student, AVD Laboratory, UT Arlington, Student AIAA Member. 3 Post Doc, AVD Laboratory, UT Arlington. 4 Associate Professor, Director of AVD Laboratory, UT Arlington, AIAA Member
F
2
Envision the next-generation rocket scientist, a human aerospace design engineer, supported by an artificial
intelligence (AI) assistant that correctly and comprehensively supports the design of aerospace vehicles or space
transportation architectures with respect to highly multi-disciplinary domains stemming from the marketplace, the
environment, politics, economics, and technology founded in an ever-changing real world. This is the overall objective
of this research.
Presented in this document is a proposition for the development of a conceptual design phase artificial intelligence
design and decision aid tool to augment and enhance the efficiency of the design engineer and decision-maker alike.
This chapter presents the contextual background of the research and the research objective. This is followed by the
discussion of intelligence and the identification of a research endeavor to develop an artificial intelligence (AI)
solution concept to assist the conceptual technology forecasting or future projects team member. This paper concludes
with the presentation of the current state of the research, specifically the test of aerospace reentry vehicle synthesis
via a neural network.
III. Problem Identification
An aerospace vehicle is a product of a
specific sequence of development and
testing; this sequence of product
development is referred to as the product
life cycle (PLC). Classically, it can be
broken down into six phases. The phases
are: (1) Conceptual Design (CD), (2)
Preliminary Design (PD), (3) Detailed
Design (DD), (4) Flight Test, Certification,
and Manufacturing (FT/C/M), (5)
Operation, and (6) Incident/Accident
Investigation (I/AI). The CD, PD, and DD
phases are considered the general design
phases. Each phase represents different
inputs, tasks, and outputs—completion of
which occurs with different toolsets and
toolset fidelity. The design knowledge and
freedom available, and the design change
cost, those attributes characterize each
phase. The result is that the CD phase is the initial critical phase responsible towards identifying the solution concept
towards ensuring project success.
1. Knowledge & Design Freedom
Knowledge and design freedom during the PLC phases are not constant. Knowledge and design freedom are
inversely related. As depicted in Figure 1, the knowledge available is minimal initially during the CD phase and
increases nonlinearly through the PLC phases. The design freedom is exactly the opposite. The maximum design
freedom available coincides with the point of minimum knowledge and decreases rapidly through the PLC phases.
2. Cost
The cost for significant design changes increases with the PLC phase. Nicolai states, “…the cost of making a
design change is small during conceptual design but is extremely large during detail design.” [1] This nature is
reflected in Figure 1. In order to minimize potential cost, it is imperative that the correct design be selected early
during the design process, which principally occurs during the CD phase
3. CD Phase
The CD phase is the phase in which the general design is selected. As postulated by Coleman, “… [t]he
fundamental objective of this conceptual design phase is to satisfy the designer and decision maker that the selected
concept is worthy of preliminary design continuation.” [2] Similarly, Torenbeek reflects that “… [t]he object of this
conceptual design phase is to investigate the viability of the project and to obtain a first impression of its most
important characteristics.” [3] The CD phase analysis results in the determination of the primary vehicle concept,
configuration, and key design parameters. [4] By the end of the CD phase, approximately 80% of the vehicle
configuration is established. [5] The inverse nature of the knowledge available and design freedom, the cost of major
design change, and the purpose of the CD phase, makes it the critical phase of the overall design process.
Figure 1. PLC with design freedom, knowledge, and cost of
change represented
CD PD DD FT,C,M O I/AI
Design Freedom
Knowledge
Cost of Change
Design
Freeze
3
A. Difficulties Affecting the CD Phase
The criticality of the CD phase does not merit the exception of a tendency to difficulty mitigation or issue
occurrence. The CD phase exhibits several design difficulty issues, two critical ones are: (1) design variable abundance
and (2) design proficiency and multidisciplinary integration decrease. (Note however that these issues are not
necessarily unique to the CD phase.) Each is addressed below.
1. Design Variable Option Abundance
The CD phase is characterized by a design freedom that translates directly into abundant design variable options
and large datasets that require assistance in interpretation and handling. The synthesis-design process is at a “…stage
of the design, [where] every parameter of design may correspond to a fairy large set of options.” [6] Figure 2 illustrates
the technical architecture level combination of mission-hardware-technology elements that come together to develop
an aerospace system. As it can be seen from the figure, the total number of possible combinations (theoretically)—
when only varying the mission and vehicle level options—increases rapidly to approximately two million distinct
vehicle concepts. It must be noted here that this estimate does not even consider disciplinary specific parametric
variations which if included would balloon the design options infinitely. Unsuspectingly, a large set of design
parameters, each parameter corresponding to a large set of option combinations, translates to large datasets. The result
is the daunting task to understand the significance of a variable among a multitude and make sense of massive
quantities of data. This quantity is too significant for an individual to assess and comprehended all variables and
subsequent combinations in this multidisciplinary cause-effect maze. Hence, it is critical that a physics-based AI
capability be developed that can parametrically trace and evaluate parametric design combinations, learn to identify
the best combinations, and ultimately augment the engineer through an AI-based learning environment. In short, the
human designer needs AI assistance.
Figure 2. Illustration of orbital reentry vehicle design variables exponential increase to unsurmountable
numbers
2. Design Proficiency
The CD phase is unique in the sense that it requires idea generation and therefore creativity and experience.
However, an interesting trend has developed; the project exposure an engineer experiences is decreasing significantly.
4
Half a century ago, an engineer could expect to work on a dozen or more projects. Today, they may be lucky to see
the completion of more than one. [7] The result of this phenomenon is the reduction in design exposure, design
experience, and subsequently design knowledge. All of which are invaluable to a designer. This illustrates a situation
necessitating a system of standardized knowledge retention, transfer, and expression.
Furthermore, there is evidence for a decreasing trend in tool integration while tool accuracy has simultaneously
been increasing. Oza points out “…that qualitatively there is a noticeable change in product development vehicle
sizing tool capability that spans the major eras of technology change.” [8] He has observed that tool accuracy is
outperforming the capture of multidisciplinary effects. This is illustrated in the figure below. Although accuracy is
important, it is problematic if the toolsets and mindset of engineers and forecasters loses the ability or understanding
of the significance of the design multivariate observability, testability, and understandability.
Figure 3. Forecasting capability: tool integration vs. tool accuracy [8]
B. Synthesis in Aerospace
Recalling that the objective is to arrive at the next generation toolset for and to enhance the engineer and forecaster,
a consideration of the progression to the current synthesis toolsets is considered. It is necessary to first understand
current approaches and evaluate how they can be improved and advanced into an intelligent frame work. Furthermore,
a new classification scheme is established. A summary of the synthesis review is illustrated in Figure 4.
Chudoba [4] provides a historical review of flight vehicle design synthesis systems and tracks the evolution in
design methodologies from the legacy textbook synthesis processes to the modern-day computerized synthesis
systems. A hierarchy of five generations of synthesis systems is defined based on the level of increasing proficiency
at integrating multi-disciplinary effects, see Table 1. The classification scheme selected distinguishes the multitude of
vehicle analysis and synthesis approaches according to their modeling complexity, thereby expressing their limitations
and potential. The first four generations of synthesis systems address modeling-complexity evolution of design
approaches from 1905 to present day capability, highlighting primary characteristics of each class.
The transition from Class II to Class III represents the first use of computer automation in the design environment.
These early design methodologies are found to focus on the selected discipline-specific analysis but lack the
multidisciplinary integration that is later implemented manually. Lovell comments that, “…initial computer
applications were confined to aspects of structural analysis and wing design. There was some resistance to the use of
computers in initial project design because of the complex decision-making process involved. However, they enabled
more detailed analyses to be made and hence allowed a greater range of carpet plots with additional overlays to be
prepared to show the effects of configuration variables on performance” [9]
Class IV synthesis systems are identified to involve multidisciplinary integration of disciplinary analysis but are
limited in application to a single-point design optimization and mostly applicable to one specific vehicle configuration.
The majority of synthesis systems up to Class IV are applicable only for subsonic and supersonic aircrafts while only
45
55
65
75
85
95
105
1960 1970 1980 1990 2000 2010 2020
Mu
ltid
iscip
lin
ary
To
ol
Inte
gra
tio
n C
ap
ab
ilit
y
Mu
ltid
iscip
lin
ary
To
ol
Accu
racy
Tool Integration
Tool Accuracy
Apollo Era STS Era SSTO Era TSTO Era ISS / Tourism Era SLS Era
Do
D,
Ind
ust
ry,
NA
SA
, e
tc
KEY RESULT: The cyclical nature of aerospace has lead to an environment where tool accuracy is outperforming the capture of multidisciplinary effects
5
a select few address the hypersonic vehicle class. Synthesis systems like Czysz’s Hypersonic Convergence [10] and
PrADO Hy. [11] are identified as significant methodology implementations of Class IV type systems.
Table 1. Five generations of evolution of CD Synthesis approach by Chudoba [4]
The assessment leads Chudoba to define the requirements for the next generation of Class V - Generic Synthesis
Capability, which is identified as a design process rather than a design tool. In this regard, the focus here is on
developing the capability over its application. The primary emphasis in this class is on the development of modular
and dedicated disciplinary methods libraries and their integration into a central multi-disciplinary synthesis
architecture.
In continuation of Chudoba’s review of synthesis approaches, Huang [12], Coleman [13], Gonzalez [14],
Omoragbon [15] and Oza [16] have conducted additional surveys of existing aerospace vehicle synthesis approaches.
These reviews cover a total of 126 synthesis approaches which include legacy textbook design synthesis
methodologies and modern-day computerized synthesis systems.
Based on these reviews, the following conclusions provide an overview summary of the existing capability and
major drawbacks of the traditional and current design methodologies (these methodologies fall under Class IV
according to Chudoba’s classification, see Table-1):
1. The majority of the existing synthesis systems have been developed for aircraft design application. Only
selected few design synthesis systems exist that address hypersonic vehicle systems. Particularly, an
efficient and dedicated design synthesis systems for highly integrated hypersonic vehicles is still missing
that has to quantifiably forecast the mission-configuration-technology scenarios.
2. Synthesis is the primary integration capability that is the key to close (converge) the design through
iterations.
3. Synthesis system are not able to efficiently define the design solution space topography; optimization is
a preferred approach not the total picture.
4. Many design synthesis systems tend to have a common structure with different computational procedures.
However, the design methodologies of synthesis systems are not transparent. There is a lack of efficient
computerized synthesis systems and multi-disciplinary interaction at the conceptual design level.
5. Existing synthesis systems have been developed specifically for a particular type of application (e.g.
subsonic, supersonic, airbreather, rocket propulsion, wing body, lifting body etc.). This implies that the
many initial assumptions and methods that are hard-coded at the development stages of the synthesis
system and limit its application to only that specific. As the system is applied over time, it becomes
hindered and stagnated, limited by the initial application boundaries. There is no generic synthesis system
for the flight vehicle conceptual design that can be consistently applied to several applications and
produce a fair non-partial assessment. This inability impedes the system’s ability to assess all available
design options and provide the best design solution independent of hardware, configuration, and
technology specifications.
The final outcome of the Gonzalez [14] endeavor was the successful development of a state-of-the-art Class-V
capability, AVD-DBMS (Aerospace Vehicle Design Database Management System). The AVD-DBMS is a proven
(see examples [17] [18] [19]) Class-V platform that is an instrument to generate unique user-specified problem-
specific sizing code (traditionally represented by Class-IV) with complete method and process transparency. The
AVD-DBMS is shown to provide the flexibility to rapidly create a new sizing code specifically tailored for
independent trade execution as required by the design problem at hand. Furthermore, this allows for parallel sizing
Class Design
Definition Develop Time Characteristics
Class I Early Dawn Until 1905 Trial and error approach, experimental, no systematic
methodology
Class II Manual Design
Sequence 1905-1955
Physical design transparency, parameter studies, standard
aircraft design handbooks
Class III Computer
Automation 1955-Today
Computerization of methods, reduced design cycles, detailed
exploration of the design space, discipline-specific software
Class IV Multidisciplinary
Integration 1960-Today
Computerized design system, MDO, data sharing, centralized
design
Class V Generic Design Future
Generation
Configuration independent, sophisticated design, true inverse
design capability, Knowledgebase systems
6
studies, thus enabling designers to generate a vast number of converged solutions and identify the wider solution
space. This capability and approach allows the designer to explore the complete design solution space and
parametrically compare distinct design options consistently.
The current best practice approach to synthesis is modularity as represented by the AVD-DBMS, the next step in
synthesis is AI integration. Many approaches to incorporate a type of AI or machine learning techniques has been
done. Common uses include expert systems, evolutionary algorithm optimization, and knowledge based systems.
However, as in the case of the Class IV and earlier approaches, the systems are still problem specific. The next
generational system is a AI assistant that can augment the engineer and employee dynamically the systems describe
in this section.
Figure 4. Synthesis System’s Reviews by AVD
C. Summary and Outlook
Up to this point, a consideration of the product design cycle, the CD phase significance and inherent difficulties,
and a synthesis summary review have been discussed. From this discussion, is evident that the early phase engineer
and forecaster require adequate assistance and augmentation through their toolsets. The current state of toolsets shows
the need for adaptability, transparency, and modularity. The solution to which is a composable multidisciplinary
synthesis design fabricator as found in the AVD-DBMS. The next evolution in synthesis is the full integration and
development of AI. This brings forward the next issue, defining intelligence and laying out a next generation AI
assistant framework.
IV. Intelligent Assistant
In the current chapter there are three points of discussion. First, a consideration of intelligence, both human and
artificial intelligence, is provided. This is followed by the general solution concept for the AI assistant where system
capability and criteria requirements are identified. The chapter ends with a development roadmap to arrive at an AI
assistant.
A. Intelligence
In the following section, we address two primary questions. What is intelligence? What is artificial intelligence?
The intent is to establish a working definition of intelligence to guide and support the definition of an artificial
intelligence design environment and subsequent solution concept.
7
1. Human Intelligence
The subject of human intelligence and identifying (let alone quantifying) it is a large topic unto itself and as such
only a brief discussion is given here. The discussion of intelligence is ancient, going back thousands of years. For the
purposes of this document, consider only the last century. A common interpretation of intelligence is the notion of
multiple intelligences. In the late 1930s, Thurstone [20] correlated intelligence to multiple abilities, identifying nine
(verbal comprehension, reasoning, perceptual, speed, numerical ability, word fluency, associative memory, spatial
visualization). Gardner [21] similarly identified intelligence as multiple intelligences working together (Visual-
Spatial, Verbal-Linguistic, Bodily-Kinesthetic, Logical-Mathematical, Interpersonal, Musical, Intrapersonal,
Naturalistic). Many of these categories such as computer vision and natural language processing are currently reflected
in the field of AI. Intelligence has also been classified as attributes. Sternber [22] identifies three attributes of
intelligence: (1) analytical intelligence, (2) creative intelligence, (3) practical intelligence. These attributes translate
to applicable aspects as problem solving, application of past knowledge to new situations, and adaptability to a new
environment respectively. With these considerations regarding the discussed various attributes of the human
intelligence, the next section discusses artificial intelligence.
2. Artificial Intelligence
The definition of AI depends on the individual asked. In its broadest state,
artificial intelligence is the mimicking of human intelligence by a computational
means. As stated by Munakata, AI is “…the study of making computers do things
that the human needs intelligence to do.” [23] Russel further breaks the definition
down based upon thought process and reasoning, and behavior arriving at four
distinct definition categories. [24] The four definition categories are (1) systems
that think like humans, (2) systems that act like humans, (3) systems that think
rationally, and (4) systems that act rationally. Given that the CD phase is the arena
of the proposed research and synthesis system development is the objective, the
broad working definition of AI, is a system that acts rationally where a “…system
is rational if it does the ‘right thing,’ given what it knows.” [24] For the purposes
of engineering it is logical to arrive at Russel’s definition. However, the definition
needs to be expanded into an even more usable sense. Human intelligence is a
multidimensional ability to apply past knowledge and experience to adapt to a new
situation and solve a problem. If this notion of human intelligence is correlated to
rational acting as described by Russel and combined with a best practice approach
to engineering (see Figure 5), a working definition of artificial intelligence for
engineering design application is arrived at. Translating this to the engineering and
computer domain, artificial intelligence is the self-utilization of a connected
database, knowledge base, parametric process, and logic base to arrive at and adapt
to a new situation to solve a problem and derive new understanding of a topic
following the rational application of an integrated, converging, and self-assessing
process.
B. System Concept
Based on the consideration of the CD phase, it has been established that the CD phase is grossly hindered by the
problem of knowledge and proficiency degradation and retention. Based on the synthesis system review, the resulting
conclusion is that most computational toolsets are highly developed high-fidelity thus high-inertia tools that generally
require excessive source code familiarity and user time input to produce any significant modification to allow the
system to address unique applications and configurations not originally considered in the tool development. With these
conditions, a next-generation system is not only logical but urgently required. This is best stated by the following:
Currently, any design synthesis or design update depends on the designer's ideas and experience base on an
ad hoc basis. Possible approaches to technology leaps in this area include idea stimulus approaches; use of
artificial intelligence and knowledge-based systems to convert designer's judgments and rules of thumb into
algorithms; techniques for visualization of the design space; multidisciplinary optimization; and automated
synthesis or inverse engineering [25]
To that end, the objective is to develop a Class VI system.
The AI solution concept is now outlined and discussed in this section. The fundamental criteria requirements for
a best-practice conceptual design toolset have been deduced through the synthesis system’s review. These criteria act
as the merit of measure for the fundamental solution logic and define the primary characteristic attributes of the system.
We begin by identifying several key system criteria requirements and follow with a general definition and explanation
Figure 5. Design ladder
8
of the solution concept. The requirements of the system’s capability are determined by the now identified character of
intelligence and the necessities of the highly fluid and non-static nature of the conceptual design phase.
1. System Criteria Requirements
Based on the synthesis system review, a necessity of a toolset is identified that is capable of rapid turnaround with
a modular structure to adapt to new concepts and configurations, thus providing the designer with an advanced
synthesis capability to develop comprehensive solution design topographies. Such a tool would require minimal user
knowledge of the system and minimum user time investment for the output of a useable synthesis or sizing code with
awareness of the developing world of artificial intelligence and potential future application.
The system criteria requirements are determined by the now identified shortfalls of current synthesis systems and
the necessities of the highly fluid and non-static nature of the conceptual design phase. The following are identified
as the primary requirements criteria for a best-practice next-generation synthesis capability:
• Flexibility: modularity to handle any fidelity and unique concept or configuration.
• Adaptability: ability to adapt to new problems, vehicles, or configurations while maintaining
applicability and usability.
• Expandability: ability to expand the underlying framework and capability when new data, knowledge,
and processes are added
• Transparency: transparent to the user and customer (if desired) of the operation of processes and
systems, the methods, underlying knowledge and data.
• Rapidity: quick turnaround, able to adapt and keep up with a rapid environment and quick
turnaround deliverable times
• Automation: key pre-requisite to arrive at an efficient assistant.
2. System Capability Requirements
The above requirements translate into a CD phase capability (where synthesis occurs) that is critical to design
success. This next generation AI environment is tasked to provide the individuals involved access to the best tools
available. Since modern synthesis methods the aerospace industry is experiencing will continue to grow in complexity,
the AI implementation has a primary focus on time to market and cost. Since the environment will continue to become
even more computationally rich in the data-knowledge-process domains, beyond the reasoning abilities of humans,
automated AI synthesis is the primer with the following system requirements:
• knowledge generation and retention through dynamic knowledge base & data base;
• scenario based multi-disciplinary analysis (MDA) design and integration;
• self-composing architecture capability with configuration, hardware, and mission independence;
• visualization and interpretation of design space topography;
• natural language interfacing;
• rational action without human oversight;
The objective is to create a unique system not bound by a fixed analysis structure that can self-educate, self-act, and
do so with the best critical thinking process available.
3. Concept
The topic of the AI research assistant aimed at augmenting hypersonic vehicle and technology forecasting are
both areas of high impact. They necessitate advancement due to several decades of stagnation and neglect in the
aerospace domain. The current approach to design and decision support is plagued with problems, including (a)
dependency on disciplinary optimization to convince and assure without offering a multi-disciplinary guarantee of
total system convergence, (b) non-standardized approach to synthesis system assembly, and (c) machine learning and
AI being inadequately applied, frequently relegated to optimization and parameter approximation. Upon successful
completion, this study advances the standard for design through three avenues: (1) transparency (2) design approach
standardization and consistency (3) knowledge and data retention and derivation. This platform will enable
organizations to innovate rapidly but more importantly experiment at lower cost and build shorter fail-cycles. It is the
cost of experimentation that affects the frequency to experiment and it is the lack of experimentation that hinders
knowledge expansion.
The objective is the development of an Aerospace Vehicle Design AI assistant (AVD-AI) that represents a best
practice implementation of an AI augmented multi-disciplinary vehicle synthesis framework that integrates a dynamic
Data Base System (DBS), Knowledge Base System (KBS) and Parametric Process System (PPS). The AVD-AI
system is envisioned to support the decision-making process by providing an intelligent, adaptive, and parametric
9
framework for systems design, strategic planning, and technology forecasting. AVD-AI is an evolution of the currently
existing best practice aerospace design synthesis capability residing in the AVD Laboratory, the AVD-DBMS
platform.
The DBS, KBS, and PPS implementation is analogous to and emulates the human domains of memory, experience,
and logic that enables dynamic feedback learning and decision-making capabilities of organic intelligence, see Figure
6. The symbiosis of these modules uniquely allows this first-generation artificial design intelligence synthesis system
to correctly solve multi-disciplinary aerospace-related problems. This next-generation artificial-design-intelligence
assistant correctly and comprehensively supports the human decision-maker and designer of aerospace vehicles. The
resulting system’s overall aim is at reducing cost, risk, and development time spans by accelerating the design process
by augmenting the person with an intelligent computer assistant.
Figure 6. Correlation of biologic intelligence foundation blocks to computer domain
To create such a unique entity requires four critical elements. The elements are: (1) Database (DB), (2) Knowledge
Base (KB), (3) Process Base (PB), and (4) a Great Intelligence (GI). The DB, KB, and PB provide the foundation of
which the GI acts and are the common elements to many current synthesis tools. However, the implementation and
process of integration and utilization of these three elements will be unique. Elements 1-3 have been the work of five
different PhD projects within the AVD Laboratory over the last decade. Those research efforts focused on the first
three elements. The following outlines the conjunction with the fourth element.
The GI is comprised of four fundamental elements. The elements are expert system, multi-variable multi-
disciplinary optimization, Pareto Analysis, and data mining. Each element will be a different independent module.
The expert system requires the data
mining and Pareto Analysis to feed the
inference system for knowledge
deduction and induction. The Pareto
Analysis is a logic element for the system
to identify relevant design variables to
execute logic algorithms. The GI will
result in a capability to retain design
knowledge, identify key design variables,
provide an independent design capability
without human logic input, provide a
solution space identification capability,
and derive new knowledge of the system
studied all while acting independently and
rationally.
Figure 7. General AI solution concept blocks
10
C. Development Outlook
The drive towards an intelligent assistant has been organized into a. five sequence approach. Each sequence is
composed of individual tasks.
Figure 8 shows an overview of the approach. The sequence’s five sub-phases with primary objectives are defined
below.
I.1) Preparation: Establish Situational Awareness of the current state-of-the-art in the selected AI domain through
a dedicated literature review. Identify, test, and evaluate approaches for system use and integration. Identify and
formulate solution concept.
I.2) Experiment Setup: System Conditioning tasks are performed by preparing training data; fundamental system
building blocks are populated, trained, and formulated (such systems are database, knowledgebase, methods library,
and evaluation metrics).
I.3) Experiment Execution: System Development occurs here through integration of fundamental building blocks
(defined in previous step) in the AVD-DBMS platform.
I.4) Prototype Evaluation: System Execution takes place in this sub-phase where the AVD-AI capability is
validated through a synthesis case-study demonstrating impact of inclusion of prior knowledge on enhanced
performance.
At this point, research activities are in sequence 1.1. The remaining elements of this paper presents the results of
an exploratory study to evaluate the potential for neural network application to sizing approximation.
Figure 8. Phase I tasks, schedule, and milestones overview
V. Synthesis Augmentation
In the current section, the experiment setup for the machine learning approach is provided followed by the
execution results.
A. Purpose
The purpose of this experiment is to test the viability of using machine learning, specifically a neural network, to
predict sizing outputs. The objective is to identify if the approach delivers sufficiently accurate results to provide a
first estimate to a sizing routine. It is expected that such a system could be used to provide a sufficiently accurate first
guess of the independent convergence variables in order to accelerate the convergence process. Additionally, a
secondary purpose is to evaluate the accuracy of a machine learning augmentation in the sizing application. In regards
to the execution of the machine learning process, a frequent issue is that the results are obtained without a physical
understanding of how or why the results are correct or arrived at exists. The opposite is true for this situation. The
underlying physical models are sufficiently known, however, the execution process is time expensive. The goal is to
skip the known physics to arrive at a suitably accurate result.
11
B. Approach
1. Process
After generation of the test data two sets of matrices of a) input, and b) target are needed. With the help of a built-
in Fitting App within the Neural Network toolbox, the data will be ready for the training phase. Figure 9 shows the
process flow of the method implement.
DATA(Input,Taget)
Neural NetworkFitting
Validation & Test DATANetwork Architecture
Train Network Method Performance Check
Regression Check
NN Function
Figure 9. Data training flow diagram
The following steps define the process to develop a neural network function which is utilized for this study:
1) Pick the sample data from workspace.
2) Define Input and Target values.
3) Assigning Validation and Test data as fraction of the whole.
4) Assigning the number of the desired networks.
5) Choose a training method and train the data correspondingly.
6) Check the performance and the regression of the data.
7) If the desired analysis achieved, export the function code for usage.
2. Data Generation
With the help of AVD-DBMS, a total of 1200 lifting body vehicle cases have been sized for this experiment. The
overall study is broken down into three phases, a) Gather the neural network function for all the 1200 cases, b)
Randomly select 240 cases and re-train the network, c) Randomly select 60 cases and re-train the network. The inputs
to the training system consist of slenderness ratio (τ), cross-section base area shape, cargo weight, and ΔV.
Decreasing the number of training cases at each time will allow to check for the accuracy of the network study and
the final differences. If less data proves that the system keeps it accuracy at a desired range, the run time can decrease
drastically. The training methods used throughout this study are a) Levenberg-Marquardt b) Bayesian Regularization
c) Scaled Conjugate Gradient.
C. Results
Table 2 shows the comparison results of trained data of 1200 different lifting body vehicles versus the DBMS
calculated value. The purpose of this table is to check the accuracy of the method and check even if this approach is
applicable. By analyzing the data, it is calculated that Bayesian Regularization Method (b) provided the higher
accuracy for this study. Next step of this study is to reduce and eliminate the initial training data. This means, by
training the neural network system with less number of sample data, the system can generate same accuracy in it’s
results.
Table 3 shows the results using only 240 points to train the system and input five random points outside of the
trained data boundary to assess the accuracy of the trained system. It is visible that the %Error increased in comparison
to Table 2 but is still within the desirable range. Again, the Method (b) shows a better accuracy for this study. Lastly
the same study has been conducted but only with 60 points to train. Based on presented results in Table 4, it is clear
that the less training point will decrease the accuracy of the system when the input outside the training boundaries are
used. These results can now be used as initial guess into the DBMS system to help forecaster/designer to arrive at their
desired converged vehicles is faster manner.
12
Table 2. Comparison between three neural network training methods
Actual Experimental % Error
Method Splan W/Splan Splan W/Splan Splan W/Splan
a
P1 28.49 8750 31.33 8754 9.06 0.04
P2 24.04 9672 28.93 9664 16.90 -0.08
P3 51.41 6840 44.24 6842 -16.19 0.04
P4 89.76 4202 90.40 4206 0.70 0.08
P5 28.37 11424 29.61 11446 4.19 0.19
b
P1 28.49 8750 28.57 8750 0.28 0.00
P2 24.04 9672 24.28 9672 1.00 0.00
P3 51.41 6840 50.40 6839 -2.00 -0.01
P4 89.76 4202 89.41 4202 -0.40 0.00
P5 28.37 11424 28.59 11424 0.78 0.00
c
P1 28.49 8750 -119.69 8613 123.81 -1.59
P2 24.04 9672 164.46 9847 85.38 1.78
P3 51.41 6840 136.74 6238 62.41 -9.64
P4 89.76 4202 96.98 3785 7.44 -11.02
P5 28.37 11424 57.49 11834 50.65 3.46
Table 3. Accuracy comparison of 240 points with random inputs
Actual Experimental % Error
Method Splan W/Splan Splan W/Splan Splan W/Splan
a
P1 53.31 3550 52.53 3553 -1.49 0.09
P2 31.86 8223 30.15 8238 -5.68 0.19
P3 24.04 9672 24.13 9659 0.36 -0.14
P4 59.38 1812 61.09 1805 2.80 -0.43
P5 34.64 5418 34.92 5420 0.80 0.03
b
P1 53.31 3550 53.92 3558 1.13 0.24
P2 31.86 8223 30.80 8203 -3.44 -0.25
P3 24.04 9672 24.59 9623 2.24 -0.51
P4 59.38 1812 58.39 1805 -1.69 -0.40
P5 34.64 5418 35.90 5461 3.52 0.78
c
P1 53.31 3550 67.51 3778 21.03 6.04
P2 31.86 8223 -18.21 9321 274.96 11.78
P3 24.04 9672 -2.40 9630 1100.14 -0.44
P4 59.38 1812 67.04 1909 11.42 5.06
P5 34.64 5418 112.29 4844 69.15 -11.84
13
Table 4. Accuracy comparison of 60 points with random inputs
Actual Experimental % Error
Method Splan W/Splan Splan W/Splan Splan W/Splan
a
P1 53.31 3550 50.88 2960 -4.79 -19.93
P2 31.86 8223 20.58 4193 -54.79 -96.11
P3 24.04 9672 21.18 6859 -13.50 -41.02
P4 59.38 1812 59.66 1810 0.46 -0.13
P5 34.64 5418 29 3816 -20.18 -41.99
b
P1 53.31 3550 51.08 2959 -4.37 -19.97
P2 31.86 8223 22.24 4210 -43.25 -95.34
P3 24.04 9672 19.59 6805 -22.74 -42.14
P4 59.38 1812 60.77 1804 2.28 -0.44
P5 34.64 5418 27 3799 -30.33 -42.62
c
P1 53.31 3550 -213.36 2979 124.99 -19.15
P2 31.86 8223 367.52 3576 91.33 -129.91
P3 24.04 9672 46.97 7150 48.82 -35.28
P4 59.38 1812 39.96 1798 -48.59 -0.81
P5 34.64 5418 -20.44 4154 269.47 -30.44
VI. Conclusion
This document communicates the research and development of an intelligent research assistant for the engineer
and decision maker endeavor at the AVD Laboratory at the University of Texas. The objective is to derive, construct,
and test a prototype AI research design assistant for technology forecasting applied to hypervelocity systems. The
proposed system will incorporate (1) best standard approaches for synthesis (2) data and physics driven high-
dimensional systems modeling and (3) knowledge derivation and retention. Fundamentally, the system proposed is an
AI-human collaborative development, discovery, and exploratory conceptual-design synthesis system approach and
execution. The objective is to push the bounds of the current state-of-the-art to illustrate a total system technique in
which conceptual design synthesis of a product is done consistently, correctly, and with the assistance of an AI
computer assistant. This assistant will be able to retain, derive, and present solution concepts, data, and new knowledge
at the benefit of the current and future human operator. The result would be a better-informed workforce allowing for
better decisions and reduction in errors and therefore costs.
The hypersonic or hypervelocity vehicle was selected for the topic for the development of the AI assistant. Today,
a key contributor aiding the development of hypersonic flight test vehicles is effective preservation and constructive
application of the available past-to-present hypersonic knowledge-base. The AI system is needed to accelerate and
leapfrog the industry to identify disruptive vehicle configurations not identified with techniques used previously. The
hypersonic vehicle drive for fruition has been slow and arduous, hampered by cyclic interest and funding. The result
has been a slow configuration evolution over the past half-century.
This document additionally communicated the results of an exploratory study into sizing augmentation by neural
networks. The objective is to investigate the applicability of a neural network trained to predict sizing architectures
converging variable output with the goal to accelerate the convergence process through improved first guess values.
The results showed a satisfactory accurate capability. The next step in this exploratory study is to expand the study to
include incomplete data sets, randomness, and convergence time improvement.
14
VII. References
[1] L. M. Nicolai and G. Carichner, Fundamentals of Aircraft and Airship Design, vol. 1, AIAA, 2010.
[2] G. Coleman, "Aircraft Conceptual Design - An Adaptable Parametric Sizing Methodology," 2010.
[3] E. Torenbeek, "Synthesis of Subsonic Airplane Design, 1982," Delft: Springer.
[4] B. Chudoba, Stability and Control of conventional and Unconventional Aircraft Configurations: a Generic
Approach, BoD--Books on Demand, 2001.
[5] B. Chudoba and W. Heinze, "Evolution of Generic Flight Vehicle Design Synthesis," The Aeronautical Journal,
vol. 114, no. 1159, pp. 549-567, 2010.
[6] C. Krishnamoorthy and S. Rajeev, Artificial Intelligence and Expert Systems for Engineers, vol. 11, CRC press,
1996.
[7] B. Chudoba, "Managerial Implications of Generic Flight Vehicle Design Synthesis," in AIAA Paper, AIAA-2006-
1178, 44th AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, 2006.
[8] A. Oza, "A Portfolio-based Approach to Evaluate Aerospace R&D Problem Formulation Towards Parametric
Synthesis Tool Design," 2016.
[9] D. Lovell, "Some Experiences with Numerical Optimisation in Aircraft Specification and Preliminary Design
Studies," in International Congress of the Aerospace Sciences, Munich, 1980.
[10] McDonnell-Douglas, "Hypersonic Research Facilities Study. Volume II Part 2 Phase I Preliminary Studies Flight
Vehicle Synthesis," NASA-CR-114324, NASA, 1970.
[11] W. Heinze, "Ein Beitrag Zur Quantitativen Analyse Der Technischen Und Wirtschaftlichen Auslegungsgrenzen
Verschiedener Flugzeugkonzepte Fur Den Transport Grosser Nutzlasten," Ph.D. Dissertation, 1994.
[12] X. Huang, "A Prototype Computerized Synthesis Methodology for Generic Space Access Vehicle (SAV)
Conceptual Design," Ph.D. Dissertation, The University of Texas at Arlington, Arlington, TX, 2005.
[13] G. Coleman, "Aircraft Conceptual Design-an Adaptable Parametric Sizing Methodology," Ph.D. Dissertation, The
University of Texas at Arlington, Arlington, TX, 2010.
[14] L. Gonzalez, "Complex Multidisciplinary System Composition for Aerospace Vehicle Conceptual Design," Ph.D.
Dissertation, The University of Texas at Arlington, Alrington, TX, 2016.
[15] A. Omoragbon, "Complex Multidisciplinary Systems Decomposition for Aerospace Vehicle Conceptual Design and
Technology Acquisition," Ph.D. Dissertation, The University of Texas at Arlington, Arlington, TX, 2016.
[16] A. Oza, "Integration of a Portfolio-based Approach to Evaluate Aerospace R and D Problem Formulation Into a
Parametric Synthesis Tool," Ph.D. Dissertation, The University of Texas at Arlington, Arlington, TX, 2016.
[17] L. Rana, T. McCall and B. Chudoba, "Parametric Sizing Boeing X-20 DynaSoar to Gain Program Architectural
Understanding of Sierra Nevada Corporation’s Dream Chaser (AIAA 2017-5355)," in AIAA SPACE and
Astronautics Forum and Exposition, Orlando, FL, 2017.
[18] L. Rana, T. McCall, J. Haley and B. Chudoba, "Conceptual Design Solution Space Identification and Evaluation of
Orbital Lifting Reentry Vehicles based on Generic Wing-Body Configuration (AIAA 2017-5127)," in AIAA
SPACE and Astronautics Forum and Exposition, Orlando, FL, 2017.
[19] L. Rana, T. McCall, J. Haley and B. Chudoba, "Parametric Sizing Implementation for Generic Lifting-Body
Configuration Executing a Low Earth Orbit Mission (AIAA 2017-5356)," in AIAA SPACE and Astronautics
Forum and Exposition, Orlando, FL, 2017.
[20] L. L. Thurstone, Primary mental abilities., Chicago, 1938.
[21] H. Gardner, Frames of Mind: The Theory of Multiple Intelligences, Basic Books, 1993.
[22] R. J. Sternberg, Beyond IQ: A Triarchic Theory Of Human Intelligence, Cambridge University Press, 1985.
[23] T. Munakata, Fundamentals of The New Artificial Intelligence: Neural, Evolutionary, Fuzzy and More, Springer
Science & Business Media, 2008.
[24] S. Russell, P. Norvig and A. Intelligence, Artificial Intelligence A Modern Approach, 2 ed., Pearson Education Inc,
2003, p. 27.
[25] J. Blair, R. Ryan, L. Schutzenhofer and W. Humphries, "Launch Vehicle Design Process: Characterization,
Technical Integration, and Lessons Learned," 2001.